TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 24, No. 3, June 2026, pp. 945 ∼ 956 ISSN: 1693-6930, DOI: 10.12928/TELK OMNIKA.v24i3.26961 ❒ 945 A segmentation based optical character r ecognition system f or Bangla printed text Mahir Mahb ub 1,2 , Ahmedul Kabir 2 1 Department of IoT and Robotics Engineering, Bang abandhu Sheikh Mujib ur Rahman Digital Uni v ersity , Gazipur , Bangladesh 2 Institute of Information T echnology , Uni v ersity of Dhaka, Dhaka, Bangladesh Article Inf o Article history: Recei v ed Feb 5, 2025 Re vised Dec 8, 2025 Accepted Mar 29, 2026 K eyw ords: Mask ed w ord prediction Optical character recognition P attern recognition Spell check er T e xtual image se gmentation ABSTRA CT Bangla ranks as the fth most spok en language globally , catalyzing signicant interest in the de v elopment of B angla optical character recognition (OCR) sys- tems. The intricate structure of the Bangla script, including compound char - acters, modiers, and headlines, complicates the formation of w ords. This research introduces a complete OCR system pipeline for printed Bangla te xt. It emplo ys a thinning-based se gmentation approach combined with a con v olu- tional neural netw ork (CNN) to recognize Bangla fonts. Additionally , a part of speech (POS)-a w are spell check er is proposed that automatically corrects mis- spelled w ords while considering their conte xt within the sentence. W e intro- duce semi-generalized lters that a dapt to ne w fonts, addressing conjunct for - mation challenges in Bangla OCR. This e xible design allo ws for ada ptation to ne w fonts. The ResNet50 model is utilized to accurately recognize se gmented characters and modiers. W e achie v e a character se gmentation error of 3.354% and an o v erall se gmentation error of 2.332%. The ResNet50 recognition model achie v es an accurac y of 98.345%. This is an open access article under the CC BY -SA license . Corresponding A uthor: Ahmedul Kabir Institute of Information T echnology , Uni v ersity of Dhaka Suhra w ardi Udyan Rd, Dhaka 1200, Bangladesh Email: kabir@iit.du.ac.bd 1. INTR ODUCTION Optical character recognition (OCR) con v erts te xt images into machine-readable form and is no w a core component in digitizing books, archi v es, administrati v e records, and e v eryday documents. Rob ust OCR systems enable information retrie v al and natural language processing and are crucial to preserving linguistic heritage and supporting digital inclusion. Bangla is one of the most widely spok en languages in the w orld, used by hundreds of millions of speak ers, is used as the of cial or national language in Bangladesh, and is a major language in parts of India. Despite its global signicance, Bangla still lacks OCR systems as rob ust as those a v ailable for man y languages (e.g., English, German, and Spanish) and e v en for se v eral Indo-Iranian languages, such as Hindi [1] and Assamese [2]. Existing Bangla OCR often st ruggles with di v erse fonts, noisy documents, and comple x page layouts. Linguistically , Bangla is clas sied within the Indo-Iranian subgroup of the Indo-European language f amily and shares le xical and structural features with Hindi and Assamese. As a res ult, OCR studies on these related scripts of fer an important basis. Ho we v er , Bangla presents unique script-specic dif culties that limit the straightforw ard adaptation of e xisting OCR approaches. Bangla uses multiple modiers (e.g., Y a-phal ¯ a , J ournal homepage: http://journal.uad.ac.id/inde x.php/TELK OMNIKA Evaluation Warning : The document was created with Spire.PDF for Python.
946 ❒ ISSN: 1693-6930 Ra-phal ¯ a , Hr as-i-kar , and R ¯ epha ) that v ary in position and shape, and its characters and modiers frequently mer ge or connect through a common headline (aka matr a ), making character boundaries dif cult to localize. These script proper ties lead to non-tri vial se gmentation and recognition problems. F or e xample, a single Bangla w ord can span three v ertical zones: an upper zone abo v e the headline, a middle zone containing the primary body of characters, and a lo wer zone belo w the baseline, as sho wn in Figure 1. The headline typically appears as the bar with maximum height i n the horizontal projection prole, residing in the upper half of the te xt line, while the baseline is obtained by scanning the lo wer portion of the line-se gmented image. Comple x lig atures and di v erse character patterns across fonts further complicate bot h line/w ord/character se gmentation and accurate classication. Although se v eral Bangla OCR systems ha v e been proposed, persistent performance and compatibility issues especially with comple x layouts, v arying typef aces, and noisy scans highlight the need for more comprehensi v e solutions. In this w ork, we identify k e y g aps: (i) se gmentation methods that are not adapti v e and struggle with di v erse script patterns during recognition. Limited e xtensibility for ne wly observ ed character formations and emer ging typef aces/fonts and (ii) the absence of a conte xtual and automated progressi v e spell check er/corrector to rene the ra w OCR output. Our approach addresses these g aps by redesigning the se gmentation and recognition pipeline with e xtensible, lter -based components, ResNet50 recognition and by incorporating an automated spell correction module to impro v e end-to-end Bangla OCR accurac y . Figure 1. Illustration of sections for line se gmented image Headline is the connectors of characters in a w ord, baseline separates the characters and the modi ers belo w the characters, the upper slice separated by headline i s called the ”upper portion”, the part between headline and baseline is called the ”mid portion”, and the part belo w the baseline is called the lo wer portion. Se gmentation-based OCR is widely recognized for its e xibility and control. W e aim to impro v e se gmentation of printed/type written te xt to reduce linguistic comple xity . Research in Bangla OCR is lim- ited, often tar geting specic se gmentation issues, leading to inte gration problems and sometimes irreparable errors. Despite these issues, this approach allo ws customization and control, potentially more ef fe cti v e than scanning/sliding windo w methods [3], [4]. Authors [5]-[7] pioneered an initial methodology for Bangla OCR as a result of successi v e research ef forts, using eight strok e-based techniques, a ll-circle technique, and a piece wise linear scan for te xt se gmentation. Their method struggles with characters ha ving parts in the up- per portion of the headline. F or e xample, Their approach can not se gment cases lik e (R ¯ epha) with ( (T i) , (T ii) , ( Hr as-i-kar ),and ( Dir gha-i-kar )). Their methodology is limited in its ability to separate numerous combined characters such as (T i) , and (T ii) . In our proposed approach, we solv e these issues. Sattar et al. [8] suggest se gmentation of v arious characters and modiers lik e , , (T i) , (T ii) , (Oi-kar) (Ou-kar) , ( Thi ), (Thii) , and (Chandr a-bindu) , based on distincti v e identication. Their ap- proach lacks e xplicit solutions for the cases mer ged in the upper portion of headline. Similarly , Akter et al. [9] use a scanning-based method with constraints. It mak es both approaches font-constrained and ine xible for other Bangla fonts. Our approach is more e xible, supporting lter e xtensions for easy adaptation to ne w fonts. Ro wnak et al. [10] propose a curv ed scanning method for character se gmentation to maintain character struc- ture. The y use w ord headline and line headline zones as the tw o form of headline zones, whic h help separate numerical characters. Ho we v er , the y sometimes f ail to se gment and , es p e cially when mer ged with others in the upper portion of the headline. Thus, this method adds noise and is limited to static fonts. Zahan et al . [11] introduce a tw o-zone strate gy using connected component analysis and a chain approximation algorithm with eight connecti vity to separate o v erlapping characters. Ho we v er , it also struggles with character o v erlap and mer ging issues with the headline. Maung et al . [12] present an OCR pipeline for handwritten Bangla that TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 3, June 2026: 945–956 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 947 addresses script-specic challenges, including modiers, compound characters , and diacritics. Their approach combines you only look once (Y OLO) for character detection with Ef cientNet-B4 for character recognition, achie ving accuracies of 93.87% for grapheme roots and e xceeding 98% for both v o wel and consonant diacrit- ics. A subsequent W ord2V ec-based spelling correction module further reduces the character error rate (CER), enhancing o v erall system performance. Some researchers used approaches based on support v ector machine (SVM), quadratic classier , and adv anced neural netw ork, which ha v e been sho wn to be promising for their automatic feature e xtraction capability for OCR [13]-[16]. Spell check ers can be a promising solution to correct the output of the OCR system. UzZaman and Khan [17] utilize double maphon y encoding to impro v e the accurac y of the suggested misspelled w ords in Bangla. Here, a mapping rule is used that combines the edit distance with a double metaphone. Islam et al. [18] proposed method uses the dictionary search, preceded by the suggestion of w ords based on the edit distance algorithm. The y calculate t he accurac y using the fractional accurac y method. In their other w ork, Islam et al. [19] inte grate the usage of the n-gram language model with the edit distance. Bijo y et al. [20] proposed a ne w detector -puricator -corrector frame w ork, detector–puricator–corrector spell (DPCSpell), based on denoising transformers, and present a method for the creation of lar ge-scale corpus. Erroneous characters are detected and mask ed by the detector netw ork, and the mask ed output is then further puried by the puricator . The corrector module is responsible for creating corrections. A ne w bidirectional encoder representations from transformers (BER T)-based spell check er , progressi v e stacking, has been de v eloped to impro v e training and correcti v e processes for languages such as Bangla [21]. The technique, which uses a con v olutional neural netw ork (CNN) sub-model, could reduce training time. The y tested on a 6,300-w ord dataset with spelling errors. Therefore there is no such auto-progressi v e spell check er for bangla. W e propose se v eral impro v ements to solv e the e xisting issues while ensuring the future scope for the e xtension of ne wly found patterns in characters and ne w typef aces/fonts. Our proposed s e gmentation strate gy is easy to follo w and sho ws se v eral impro v ements to pre vious se gmentation issues. W e are presenting an OCR system that does not ha v e an y compatibility issues related to se gmentation and is open for e xtension to accommodate ne w fonts. Then we feed the se gmented characters to the CNN model to recognize them while preserving the locality of the te xt position in our te xtual output. Finally , we feed the output to the proposed spelling check er model, which automat ically corrects spelling. This proposed bangla spelling check er is conte xtual, rst auto-progressi v e, and less sensiti v e to pre vious spelling errors. 2. METHOD This section pro vides an e xtensi v e elucidation of the methodology emplo yed in our research, with comprehensi v e details presented in the subsequent subsections. 2.1. Pr epr ocessing OCR preprocessing encompasses se v eral k e y steps: image binarization, sk e w correction, noise re d uc - tion, and orientation detection. W e emplo yed otsu’ s m ethod for binarization. F or sk e w correction, the approach outlined in [22] w as used. Document noise, such as salt and background interference that commonly arises from the age or quality of the document, w as mitig ated by median ltering. In addit ion, contrast-limited adapti v e histogram equalization (CLAHE) w as implemented to enhance local image details, and intensity le v els were adjusted using sigmoid correction. 2.2. Segmentation Here, we dene the methodology used to se gment the lines, w ords, and characters within the te xt images. The subsequent subsections pro vide a comprehensi v e e xplanation of these processes. 2.2.1. Line detection and segmentation T o determine the orientation of an image, we implement both horizontal and v ertical scanning tech- niques to tally the black pix els present for both ro ws and columns. This pix el count f acilitates the generation of a histogram, which is instrumental in line detection. Figure 2 illustrates an e xample of such a histogram deri v ed from the scanning proces s. Horizontal scanning of the images is emplo yed to identify lines. F ollo wing the separation of these lines, a thinning algorithm is emplo yed which is represented in Figure 3. A se gmentation based optical c har acter r eco gnition system for Bangla printed te xt (Mahir Mahb ub) Evaluation Warning : The document was created with Spire.PDF for Python.
948 ❒ ISSN: 1693-6930 Figure 2. The histogram in the v ertical and horizontal orientations sho ws the count of black pix els for each direction respecti v ely . The dotted lines indicate the separation between the lines of te xt Figure 3. The separated te xt line is thinned using sk eleton-based thinning approach 2.2.2. W ord segmentation In se v eral w ord se gmentation techniques [5], [6], a w ord is identied as a single connected component, using v ertical histogram scanning to separate w ords. Although some w ords, such as (Kha ga) , lack internal connection to the headline. T o address this, we manually e xtended the horizontal border length of the connected component by 10% to its v ertical length. W e did it because our res ults indicate that the intercharacter spacing is usually 6-10%, while the interw ord spacing is 14-25% of the w ord’ s v ertical le ng t h in standard te xt. Then lled black rectangles are superimposed on the modied border of the connected components in a cop y of the image. Then we recalculate the border of the connected component. Finally , we e xtract the areas from the main image using recalculated borders, which result in images of w ords. The outcome of the method can be seen in Figure 4. T w o separate components ( , and ) sho wed in Figure 4(a), of the w ord ( ¯ Ag ¯ e) are mer ged into a single component using the lled rectangle in Figure 4(b). (a) (b) Figure 4. An e xample of w ord se gmentation by the lled rectangle: (a) connected component: the tw o w ords are se gmented into three connected components. Those components are sho wn inside the rectangles and (b) e xtended and lled the horizontal length of the rectangle co v ering the w ord is e xtended and lled. Therefore the sub-section of the w ords are mer ged and separated as a single w ord 2.2.3. Character segmentation W e introduce common search patterns through what we call lters to se gment the characters and modiers. W e designed the lters in such a w ay that the lters do not o v erlap with each other when searching. The lters in Figure 5 remo v e the headline from the thinned image. Here, C P denotes the start of the pix el from which we start searching. P and N denote positi v e and ne g ati v e character pix el. The pix el colors for C P and P are black. P atterns matching the lters in Figures 5(a) to (e) indicate that there is only a headline. Thus, we remo v e the black pix els of C P and P . Filt ers are crucial because the headline after thinning is not al w ays a straight line, as c an be seen in Figure 6. The small rectangles in Figure 6 re v eal those v ariations in the structure of the headline. Thus these lters help identify the headline so that the y can be remo v ed. (a) (b) (c) (d) (e) Figure 5. Filters for headline remo v al: (a) lter 1, (b) lter 2, (c) lter 3, (d) lter 4, and (e) lter 5 TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 3, June 2026: 945–956 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 949 Figure 6. V ariations in the headline appear follo wing the thinning process, as indicated within the gray rectangles Our approach ef fecti v ely resolv es mer ging issues, such as touching and intersecting, as sho wn in Figure 7. W e use lters to identify and separate characters and modiers lik e (R ¯ epha) , , , , and upper portions of , , , , and as sho wn in Figure 7(a) and 7(b) . In Figures 8(a) to (e), the lters in Figures 8(a) to (c) are used to search in the image. Upon identifying these search patterns with the lters, we search for single ”up”, ”up-right”, and ”right” pix els from the current pix el accordingly . At rst we search in the straight ”up” direction. If the e xpected pix el is not found, the search mo v es to the up-right and then right. W e do this single pix el search because sometimes the pix els of the thinned headline res ide one pix el ”up” from the a v erage headline. It usuall y happens when characters join with the headline. So, the starting point is adjusted to the ne w positi v e search point upon matching this search pattern. (a) (b) Figure 7. Examples of character mer ging issues: (a) with upper portion of and (b) with upper portion of (a) (b) (c) (d) (e) Figure 8. Filter for detecting upper modiers using dif ferent search and erasing strate gies: (a) lter 1, (b) lter 2, (c) lter 3, (d) lter 4, and (e) lter 5 Then we search using depth-rst search (DFS) from the identied headline pix el. When a horizontal pattern or an intersecting point is found with at lea st a six-pix el count during the DFS search, we dra w a white lled rectangle that separates the mer ging characters ( and (Rpha) ), as sho wn in Figure 7(a). A (R ¯ epha) possibly mer ges with a headline when matched to the lter in Figure 6 which we pre viously used for adjustment of headline pix el. W e replace C P and P with N , then apply DFS ag ain using the ”up” and ”up-left” patterns. The intersection point is disco v ered by nding at least three positi v e adjacent joint pix els in the upper portion of the headline. It separates (R ¯ epha) . Lik ely , , and upper portion of , , , , e xist here when the matched pattern with the lter in Figure 8(d). Preceeded by a pattern corresponding with an y of the Figures 8(a)-8(c). Therefore, we dra w a white line, which is 20% of the width (aka. v ertical length) of the line-se gmented picture, just one pix el up from the C P pix el at the backw ard horizontal direction. This white line separates from the headline without an y deformation. Ag ain, similar to the pre vious one, we conduct a DFS based on an ”up”, ”up-left”, ”left” pattern. W e remo v e the nal positi v e search pix el after locating the intersection point. It s eparates characters/modiers to pre v ent mer ging. The whole process is described in Algorithm 1. Our method splits , , , , into tw o parts, as sho wn in Figure 9. Se gmented characters are collected on the basis of their centroids. From the centroid, the borders of the characters are calculated. Figure 9. An e xample of fully se gmented te xt. The images are separated into lines. Then the lines are separated into w ords. The w ords are separated into characters and modiers A se gmentation based optical c har acter r eco gnition system for Bangla printed te xt (Mahir Mahb ub) Evaluation Warning : The document was created with Spire.PDF for Python.
950 ❒ ISSN: 1693-6930 Algorithm 1 Separate characters and modiers from OCR image 1: pr ocedur e S E P A R A T I N G C H A R A C T E R S A N D M O D I FI R E S ( imag eAr r ay , l ineR ow I ndex ) 2: lineArray = & I mag eAr r ay [ LineR ow Ar r ay ] 3: lineLength = l eng th ( l ineAr r ay ) 4: f or i = 0 to l ineLeng th − 1 do 5: lterT ype, horizontalCount, erasingInde x = F indS eper ator F il ter ( C P = & LineAr r ay [ i ]) 6: if f il ter T y pe in [ a, b, c ] then 7: if hor iz ontal C ount ≥ 6 then 8: E r a s eB y R ectang l e ( r e ctang l eS tar tI ndex = r ectang l eI ndex ) 9: end if 10: innerFilterT ype, , erasingInde x = F indS e p er ator F il ter ( C P = & LineAr r ay [ i ]) 11: if inner F il ter T y pe == d then 12: E r aseB y B ack w ar dLine ( index = er asi ng I ndex ) 13: isMer geF ound, mer gerPix el = S ear chF or M er g e ( C P = & LineAr r ay [ i ]) 14: if isM er g eF o un d then 15: E r asing M er g er P ixel s ( mer g er P ixel = mer g er P ixel ) 16: end if 17: end if 18: else if f il ter T y pe == e then 19: PInde x es = GetP I ndexes ( C P = & LineAr r ay [ i ]) 20: isMer geF ound, mer gerPix el = S ear chF or M er g e ( C P = & LineAr r ay [ i ]) 21: if isM er g eF ound then 22: E r asing M er g er P ixel s ( mer g er P i xel = mer g er P i xel ) 23: end if 24: E r asing P ositiv eP ixel s ( C P = i, P I ndexes = P I ndexes ) 25: end if 26: end f or 27: end pr ocedur e 2.3. Recognition and post pr ocessing Our proposed model is a st andard ResNet50 CNN (pre-trained with imagenet) and tw o a d di tional feed-forw ard layers for character recognition. The model is trained with 127 classes with 10 fold cross v ali- dation. Each class is constrained to 451 labeled instances of ima ges because of the rare occurrence of some characters/modiers. W e ha v e e xperimented using other CNN models. Ev ery class has W e mer ge character output into w ords and lines. Post-processing corrects cert ain OCR generated errors. W e handle the joining character separately by mer ging it with the pre vious character , bypassing the recognizer . After headline erasing, (the letter T a ) and (the number 3) both become . Therefore we check for se v eral cases to determine if it is or . If the i th v alue in w ord is , the cases are as follo ws: If the ( i − 1) th or ( i + 1) th v alue is a number , the i th v alue is . If the ( i + 1) th v alue is and the ( i − 1) th v alue is not a number or the ( i − 1) th v alue does not e xist, the i th v alue is . Otherwise, the i th v alue is . Only unique cases in ends with . In addition, erasing the headline resulted in the se gmentation of , , , , into tw o parts. Among the se gmentation of , , can result in similarity to other characters. ’ ’ becomes ’ ’ and ’ ’, ’ ’ becomes ’ ’ and ’ ’, ’ ’ becomes ’ ’ and ’ ’. So, when these characters are found in the ith position, we check for ’ ’ in the i − 1 th position and tak e the decision accordingly . 2.4. Spell check er The spell check er we propose processes te xt obtained via OCR and outputs te xt with corrected spelling. The architecture of the spell check er is detailed in Figure 10 and is further elaborated upon in subsequent sec- tions. Our approach utilizes a BER T , which emplo ys mask ed w ord prediction to infer w ords by considering the surrounding conte xt. This mechanism generates a list of lik ely w ords to correct spelling errors. The BER T model is a base-uncased model, featuring 12 layers, 768 hidden units, 12 attention heads, and a total of 110 million parameters. The pri mary training corpus for the BER T model is sourced from Bangla ne wspaper articles, including Daily Prothom Alo,Daily Kalerk ontho, Daily Ittef aq and Bangla W ikipedia, constituting 65.36% of the dataset, with the remaining 34.64% deri v ed from Bangla W ikipedia . Data acquisition w as performed through web cra wling and manual collection between 2018 and 2020, resulting in a corpus comprising 10,340,273 sentences and 8,928,472 unique w ords. In T able 1 the data source is described using its proportional v alue from dif ferent sources. TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 3, June 2026: 945–956 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 951 Figure 10. The main components of the proposed spell check er are sho wn inside the dotted area. The spell check er tak es the sentence as input and produces the sentence with the corrected spelling of w ords T able 1. Details of the dataset used for corpus b uilding Data source Bangla ne wspaper Bangla W ikipedia Proportion(%) 65.36 34.64 2.4.1. Err or detection W e address spelli ng errors in the OCR output, rst by identifying named entities using an mBER T model (https://huggingf ace.co/sagorsark er/mbert-beng ali-ner). Non-named entities under go a dictionary lookup within our 451,742-w ord database. W ords not found and not numerical are agged as potential misspellings for further analysis. T o correct the spelling, we took a tw o-step approach. Firstly , we took the prediction for the probable w ord list from the mask ed w ord prediction model for probably misspelled w ords. The predicted w ords t h e n under go a correction process based on Le v enshtein distance (LD). LD calculates the minimum edits needed for the con v ersion of w ord s 1 to w ord s 2 . Let I represent the current location in the input string s 1 and J represent the current location in the output string s 2 . If the probable misspelled wor d (mw) length is less than 4, we consider the maximum edit distance (ml) of 2 indicated by ml to select a correctly spelled candidate w ord. If the mw or predicted w ord( pw ) has comple x characters (tw o characters joined by ” ”), the v alue of ml is e xtended by 1. F or the v alue of the length of mw > = 4, the v alue of ml is half the length of mw . Here, the selection of minimum edit distance is based on analysis and e xperiment. When predicted w ords from the mask ed model predicted w ord list fulll the criterion described abo v e, the w ord from the li st with the minimum edit distance is considered as the correctly spelled w ord and replaces the probable misspelled w ord. The model can also suggest the most probable list of correctly spelled w ords. 3. RESUL TS AND DISCUSSION Here, we discuss the result of our se gmentation process. Since our lters are currently tailored to the specic fonts used in the Daily Prothom Alo and Somokal ne wspapers, it is important to compare ho w well the method generalizes across dif ferent typef aces. The se gmentation performance for the dif ferent fonts is sho wn in T able 2. Our proposed approach has se gmented printed te xt images into 123,422 characters/modiers. The se gmentation performance for the dif ferent fonts is sho wn in T able 2. Our proposed approach has se gmented printed te xt images into 123,422 characters/modiers. The majority of characters, such as , which are occasionally aligned abo v e or with the headline, are responsible for the character -le v el se gmentation problem. F or the Daily Prothom Alo and Somokal typef ace, an additional sk e w correction of the headline after the thinning procedure resulted in a 2.26% reduction of error . In our proposed approach, we sol v e man y mer ging-related se gmentation issues. T able 2. Comparati v e result of character se gmentation for dif ferent ne wspaper fonts F ont source Character se g. err(%) T otal se g. character Prothom Alo 2.80 63,359 Somokal 4.45 32,112 Jug antor 11.38 27,951 A se gmentation based optical c har acter r eco gnition system for Bangla printed te xt (Mahir Mahb ub) Evaluation Warning : The document was created with Spire.PDF for Python.
952 ❒ ISSN: 1693-6930 In T able 3, we compare the performance of character -le v el se gmentation across dif ferent ne wspaper typef aces. This comparison is important because it re v eals certain widely used ne wspaper fonts, especially with comple x lig atures and closely spaced glyphs. It sho ws that e xtension to a ne w font f ace is important for future generalization, which scope we opened through the ltering-based approach. In our method, the a v erage se gmentation error rate is 0.22% on all typef aces. T able 3. Result of se gmentation of some mer ging issues and characters/modiers. Most of the issues arise due to the mer ging of modiers and characters Issues Character se g. err(%) Mer ging with ( Hr as-i-kar ) 0.95 Mer ging with ( Dir gha-i-kar ) 1.42 ( Hr as-i-kar ) 0.21 ( Dir gha-i-kar ) 0.13 ( Oi-kar ) 0.02 ( Ou-kar ) 0.23 ( Chandr a-bindu ) 0.48 Mer ging with ( R ¯ epha ) 0.03 ( Y a-phal ¯ a ) 0.41 ( Ra-phal ¯ a ) 6.31 ( ga ) 8.09 ( tha ) 7.23 T able 4 pro vides a direct quantitati v e comparison between our proposed method and e xisting Bangla OCR systems under the same e xperimental conditions and font settings (Prothom Alo and Somokal). This comparison i s crucial to demonstrate that performance g ains in character se gmentation and recognition are not anecdotal, b ut are consistently achie v ed despite the limited and highl y imbalanced 127-class data set. Each class is constrained to 451 labeled instances of i mages because of the rare occurrence of some characters/mod- iers. Our approach has an error rate of only 3.35%, which is signicantly lo wer than that of the pre vious methods.Gi v en the scarcity of complete OCR approaches in the literature, the nal results are compared e x- clusi v ely with studies that emplo y full OCR methods. In addition, there is a limited standard benchmark for comparing the performance of OCR [23], [10]. W e ha v e e xperimented with dif ferent CNN models for recognizing characters/modiers, as sho wn in T able 5. The ResNet50 models gi v es the best outcome compared to other e v aluated CNN models with a accurac y of 98.345% and an F1 score of 96.854%. W e e v aluate the performance of the spell check er on the Bangla BER T mask ed w ord model described abo v e. The metrics used to e v aluate the models are described: - TP (true positi v e): did not change the correct w ord/total correct w ord - FN (f alse ne g ati v e): change the correct w ord incorrectly/total correct w ord - FP (f alse positi v e): did not change the incorrect w ord (Mark incorrect as correct)/total incorrect w ord. - TN (true ne g ati v e): change the incorrect w ord correctly/total incorrect w ord - TN PLUS: change the incorrect w ord incorrectly/total incorrect w ord T able 4. Comparati v e analysis of the traditional se gmentation-based Bangla OCR systems Literature Character se g. error (%) Final outcome error (%) Our proposed approach 3.35 2.33 Cho wdhury and P al [7] 39.87 44.53 Mahmud et al. [24] 34.22 - Sattar et al. [8] 36.99 - Akter et al. [9] 42.66 - Omee et al. [23] 28.60 30.65 Ro wnak et al . [10] 23.54 - Zahan et al. [11] 28.99 - TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 3, June 2026: 945–956 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control ❒ 953 T able 5. The result of the dif ferent CNN model on labeled se gmented bangla characters/modiers Model name Accurac y (%) F1 ResNet50 98.345 96.854 V GG-16 86.433 84.121 Inception V2 89.992 85.471 ANN(10 Con v olutional Layers) 67.443 59.328 The dataset utilized for the e v aluation of spell check er performance is constructed in accordance wi th the methodology described in the literature by Sif at et al. [25]. There are 10244 misspelled w ords and 52334 correctly spelled w ords in the dataset. W e e xpect to ha v e the v alue of TP , and TN as lar ge as possible and others as lo w as possible. W e e xperimented with v arious maximum edit distance values (ml) . W e dened and check ed some conditions to decide when we should change the spelling of w ords based on the LD. The conditions are listed: - C1: ml = (probable misspell w ord aka mw’ s length(length(mw)))/2 - C2: ml = length(mw)/2 if length(mw) > 4 else ml = 2 - C3: ml = length(mw)/2 if length(mw) > 6 else ml = 2 - C4: ml = length(mw)/2 if length(mw) > 6 else ml = 3 In T able 6, C2 sho ws better results than other conditions that ha v e already been mentioned in the abo v e proposed method. Thus we select C2 as our para meterized v al ue for ml . Although C2 sho ws 8.33% in TP PLUS, the v al ue of TN is 50.17% which is signicantly good. C1 performs better in detecting the correct w ords in sentences. Finally , in T able 7, we sho w an e x a mple to demonstrate the step-by-step process of our OCR. It sho ws the dif ferent outputs of the OCR through its w ork c ycle. T able 6. The outcomes of our spell check er using Bangla BER T under v arious conditions Condition TP FN FP TN TN PLUS C1 0.983 0.016 0.677 0.320 0.001 C2 0.978 0.021 0.415 0.501 0.083 C3 0.977 0.022 0.553 0.441 0.005 C4 0.962 0.037 0.649 0.201 0.149 T able 7. Detailed vie w of the pipeline’ s processes using an e xample Step Outcome Printed e xample image Line se gmented thinned image After headline erasing Output from recognition model Output from spell check er So among the tested fonts, the proposed method achie v ed the lo west se gmentation error with P rothom Alo (2.80%), while fonts lik e Jug antor sa w higher error rates due to the l ters are not optimized for them. Compared to pre vious systems with error rates, our approach reduced character -le v el errors to 3.35% and nal te xt errors to 2.33%. The part of speech (POS) sentiti v e autoprogressi v e and conte xtual BER T -based spell check er further impro v ed recognition outcomes, particularly under condition C2. A se gmentation based optical c har acter r eco gnition system for Bangla printed te xt (Mahir Mahb ub) Evaluation Warning : The document was created with Spire.PDF for Python.
954 ❒ ISSN: 1693-6930 4. CONCLUSION In this w ork, we ha v e addressed the problem of digitizing Bangla printed images through a com- plete OCR pipeline, with particular emphasis on character se gmentation, recognition, and post-processing. W e proposed a modular , e xtendable lter -based se gmentation method that mitig ates man y of the shortcom- ings found in e xisting se gmentation-dri v en Bangla OCR approaches, thus reduci ng se gmentation-related errors across v arious printed fonts. By inte grating our se gmentat ion output with ResNet50, we achie v ed competi- ti v e character -le v el recognition performance, despite the challenges posed by the comple x Bangla script and compound characters. W e proposed a conte xtual spell check er b uilt on a mask ed language model (BER T). By operating at the phrase le v el, this component enhances the rob ustness of the OCR pipeline ag ainst both recogni- tion errors and traditional spelling errors, and sho wcases the utility of modern language models in lo w-resource script OCR. Although our proposed complete pipeline does not solv e all challenges of Bangla OCR, it establishes a strong foundation and suggests se v eral promising directions. Our current system tar gets printed te xt. Extending the se gmentation, recognition, and conte xtual correction components to handle handwritten Bangla, where the v ariability is signi cantly higher , remains a signicant and impactful research direction. The performance of both recognition and BER T -based spell correction can be substantially i mpro v ed by training and e v aluating lar ger , more di v erse datasets that better co v er rare compound characters, multiple fonts, and a wider range of domains. De v eloping additional lters and inte grating them into the e xisting systems can enhance the o v erall co v erage and rob ustness of the pipeline. Also, there are a lar ge number of compound characters in Bangla. These characters appear too rarely to be incorporated into our recognition model. These characters af fect the o v erall performance of the OCR. Although our primary objecti v e w as to enhance se gmentation techniques, it is essential to ackno wledge that se v eral contemporary studies emplo ying non-se gmentation methodologies ha v e g arnered signicant recognition and demonstrated high ef cac y . In summary , our w ork presents a se gmentation-centric pipeline for Bangla printed OCR, combining classical image processing and modern deep learning with conte xtual l anguage modeling. W e e xpect that the methods and insights introduced here will f acilitate further research in rob ust Bangla OCR and, more broadly , OCR for other comple x Indic scripts. FUNDING INFORMA TION Authors state no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENTS This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Mahir Mahb ub ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Ahmedul Kabir ✓ ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting CONFLICT OF INTEREST Authors state no conict of interest. D A T A A V AILABILITY Data supporting the ndings of this study are a v ailable upon request from the corresponding author . TELK OMNIKA T elecommun Comput El Control, V ol. 24, No. 3, June 2026: 945–956 Evaluation Warning : The document was created with Spire.PDF for Python.